Framework for clustering graphs using various distances/proximity measures.
List of distances/proximity measures:
- SP-CT: Shortest Path and Commute Time (and linear combination)
- pWalk: plain Walk (a.k.a. Von Neumann diffusion kernel)
- For: Forest (a.k.a. Regularized Laplacian kernel)
- Comm: Communicability (a.k.a. Exponential diffusion kernel)
- Heat: Heat kernel (a.k.a. Laplacian exponential diffusion kernel)
- NHeat: Normalized Heat
- Walk, logFor, logComm, logHeat, logNHeat: logarithmic versions of pWalk, For, Comm, Heat, NHeat
- SCT: Sigmoid Commute Time
- SCCT: Sigmoid Corrected Commute Time
- RSP: Randomized Shortest Path
- FE: Free Energy
- PPR: Personalized PageRank
- ModifPPR: Modified Personalized PageRank
- HeatPPR: Heat Personalized PageRank
- logPPR, logModifPPR, logHeatPPR: logarithmic versions of PPR, ModifPPR, HeatPPR
List of clustering algoritms:
- Kernel k-means
- Spectral clustering
- Ward
- Wrappers for kernel k-means from kernlab, sklearn k-means
List of graph generators:
- Stochastic Block Model
List of graph samples:
- Dolphins
- EU-core
- Football
- Zachary karate
- Newsgroup
- Polbooks
If you wish to use and cite this work, please cite this earlier paper which used many of the same concepts and methods (a newer publication is in preparation):
Ivashkin and Chebotarev, "Do logarithmic proximity measures outperform plain ones in graph clustering?" International Conference on Network Analysis, Springer, 2016.